embed: hardcoded 30-minute session maxDuration aborts large-corpus embeddings (store.js:1377)

Start at dist/store.js:1377 and inspect how qmd embed currently reads QMD_EMBED_CONTEXT_SIZE and QMD_EXPAND_CONTEXT_SIZE. Make the embedding session duration configurable while preserving the 30-minute default, then run qmd embed on a corpus that exceeds 30 minutes and verify it completes without session expiry.

Summary

The generateEmbeddings LLM session in dist/store.js:1377 has a hardcoded maxDuration: 30 * 60 * 1000 (30 minutes). On large corpora that take longer than 30 min to fully embed, the session aborts mid-run via abortController, removeIncompleteEmbeddings() purges partially-embedded documents, and the user is forced into many resume runs that may never converge.

Reproduction (qmd 2.5.1, Apple Silicon CPU)

Corpus: ~6,400 markdown docs across two collections (mostly Chinese, ~66,000 chunks total after re-chunk).
Model: Qwen3-Embedding-0.6B-Q8_0 via QMD_EMBED_MODEL.

Why resume doesn't converge: doc-level completion requires all chunks for a content hash to finish in a single session before removeIncompleteEmbeddings() keeps them. Each 30-min run re-chunks the not-fully-done set, persists ~8-9k chunks, then aborts — so vectors grow but the doc-level "done" set only advances by the docs that happened to fit entirely within a 30-min window. At the observed rate (~230-323 content hashes flipped to done per run), convergence on this corpus would take ~17-24 additional runs.

Memory note

The 2.1.0 fix (#500 / #393) holds. Memory stayed bounded at ~3 GB throughout the 3h+ patched run on Apple Silicon. The 30-min cap appears to be over-defensive given that the underlying memory leak is fixed.

Workaround (used in production)

Monkey-patch dist/store.js:1377:

// before
}, { maxDuration: 30 * 60 * 1000, name: 'generateEmbeddings' });

// after
}, { maxDuration: 4 * 60 * 60 * 1000, name: 'generateEmbeddings' });

Survives until next npm install -g @tobilu/qmd.

Proposed fix

Make it env-configurable, in the spirit of QMD_EMBED_CONTEXT_SIZE / QMD_EXPAND_CONTEXT_SIZE:

maxDuration: Number(process.env.QMD_EMBED_MAX_DURATION_MS ?? 30 * 60 * 1000)

Or add a --max-duration <ms> flag on qmd embed. The 30-min default is sensible as a safety net; large corpora need an escape hatch.

Environment

• qmd 2.5.1 (installed via npm install -g @tobilu/qmd)
• Node — system default via Homebrew
• macOS, Apple Silicon
• Embedding model: hf:Qwen/Qwen3-Embedding-0.6B-GGUF/Qwen3-Embedding-0.6B-Q8_0.gguf
• Reranker: default qwen3-reranker-0.6b